1. Segment Routing

2. Source routing
IP routing is based on destination addresses (and perhaps DSCP)
but sometimes we need control over
the precise path a packet travels to its destination
For example
in DCs we need to ensure packets traverse nodes (in order)
for security we may need to avoid a particular router
policy-based routing enables overriding default routing
we may need paths with special characteristics (e.g., low delay)
IP protocols provide mechanisms called Source Routing
IPv4 source routing options (Loose SR, Strict SR)
IPv6 type 0 routing header extension (Rh0)
SR inserts sequences of router addresses into packet headers

3. Source routing example
shortest path
SR path A B C D E
Loose SR – A C D
Strict SR – A B C D E

4. Source routing is evil Yet source routing is problematic
There are several reasons why source routing is evil
complicated processing for core routers
DoS attack – attacker forces packets to traverse selected routers, thus overloading them
amplified DoS attack – attacker forces packet to oscillate between 2 selected routers
infiltration attack – attacker bypasses ACLs by forwarding through a permitted waypoint
The IETF has not yet completely deprecated it
but highly recommends that it be disabled
Core Internet routers drop packets with options
(Linux kernels no longer process Source Routing)

5. Safe policy-based routing
But without SR, how can we achieve policy based routing?
There are 2 alternatives
Software Defined Networking
SDN give the network administer full control over routing
particular flows can be configured to traverse arbitrary paths
CE with NMS, and MPLS-TE with PCE are essentially the same
But SDN
requires relatively large architectural changes
enables attacks (and plain bugs) at control plane level
Segment Routing
Segment routing is similar to Source Routing, but
the path is specified by an ingress router, not by the source host
thus blocking Source Routing attacks (unless a router is compromised)

6. Segment routing vs. SDN
In SDN the network maintains per-application/flow state
With SR forwarding instructions provided in the packet.
In SDN all the intelligence is in the centralized controller
the SDN switches are dumb, fast, and inexpensive
SR burdens the ingress LER (like PCE)
it needs to digest the IGP, prepare the label stack, ...
OpenFlow-like based SDN has a major design flaw
flows are identified by configuring matching tables
matching table logic for 1 flow may influence other flows
so even minor bugs, and certainly malicious rules
may impact services that have been running perfectly for years
Errors in Segment Routing only affects the flow itself
Both SR and SDN can coexist with conventional networking

7. MPLS-based Segment Routing
MPLS forwards packets using a simple universal paradigm
read ToS Label
look up label in LFIB
perform label stack operation (swap, push, pop) in NHLFE
forward packet according to NHLFE
In regular MPLS networks
most of the time the label stack operation is swap
pop is used by egress LERs and FRR
MPLS segment routing reuses the standard MPLS mechanism
ingress LER inserts an entire stack of labels, one per hop
each LSR pops a label revealing the next hop
MPLS SR doesn’t require LDP or RSVP-TE (but extends the IGP)

8. MPLS Segment Routing example
desired path
label = A
ingress LER
egress LER
label = B
label = C A B C
ToS
BoS 1 2 3
Ingress LER inserts label stack with 3 labels : A (ToS), B, C (BoS)
1st LSR reads A, pops label, forwards over link for A
2nd LSR reads B, pops label, forwards over link for B
3rd LSR reads C, pops label, forwards over link for C

9. Global and local segments
In Segment Routing the labels are called Segment IDs (SIDs)
in MPLS SR the SID is the 20-bit label
and in IPv6 SR (SRv6) it is a 128-bit address
There are 2 main types of SIDs :
An adjacency SID (local SID) refers to a link (port)
it has local significance (like normal MPLS labels)
only the LSR advertising it can use it with that meaning
A node SID (prefix SID, global SID) refers to a destination node
if has global significance (unique, like IP addresses)
the network forwards over the shortest path to the node
every LSR has the same entry in its LFIB
WARNING: this is a simplification

10. Label distribution The ingress LER learns nodes and adjacencies
from the Interior Gateway Protocol (e.g., OSPF or IS-IS)
Hence, it can select each node and link
to be traversed along the desired path
The source LSR can insert
(global) node SIDs (either direct or loose)
or adjacency SIDs
or combinations
But how does the source LSR know the labels
that indicates to an LSR to forward over a desired link?
Segment Routing augments the IGP with label information
(LDP is no longer needed)

11. Segments as instructions
Constructing a segment routing label stack
is similar to programming in a low-level language
Each label can be considered to be an instruction (op-code)
The ingress LER encodes the list of instructions (SIDs)
and each LSR interprets and executes one instruction
thus making the networking into a giant processor
Segment instructions can be:
Forward over link L
Go to node N using the shortest path
Apply service S

12. IPv6 extension headers The standard IPv6 header looks like this:
and by using “Next Header” one can add options
in particular, the routing extension header
VER=6
TC 8b
Flow label 20b
Payload Length 16b
Next Header 8b
Hop Limit 8b
Source Address (SA) 128 bits
Destination Address (DA) 128 bits
Next Header 8b
Header Len 8b
options + padding
options + padding
Next Header 8b
Header Len 8b
Type 8b
Segments Left 8b
optional type-specific data

13. SRv6 extension header (SRH)
SRv6 uses the routing extension header with type = 4
and multiple SRv6 segments are concatenated
Next Header identifies the type of header after the SRH
Segments Left is decremented at each segment
Last Entry = n (the last entry in the segment list)
Flags include P (protected) O (OAM) A (Alert) and H (HMAC)
Next Header 8b
Header Len 8b
Type=4 8b
Segments Left 8b
Last Entry 8b
Flags 8b
Tag
Segment[0] 128b
Segment[1] 128b
...
Segment[n] 128b
optional TLVs

14. Unified-IP-SR There is another encapsulation for SR in IP networks
RFC 7510 defines MPLS-in-UDP for IPv4 or IPv6 networks
This encapsulation may be better than RFC 4023
since it enables fine grain load balancing using ECMP for IPv4
by using the UDP port for entropy (IPv6 already has the flow label)
Unified-IP-SR exploits MPLS-in-UDP to carry MPLS SR
Routers must be capable of this new type of forwarding
and must advertise this capability in the IGP
but Unified-IP-SR can function
with a mixture of unified-IP-SR capable and legacy routers
MPLS-in-IP
MPLS-in-GRE-in-IP

15. FRR and LFA
One of the deficiencies of SDN is the lack of resilience methods
OpenFlow provides a mechanism via group tables
Segment routing enables new resilience methods
that do not require signaling
that do not require maintaining massive network state
that avoid looping
One such method is
Topology Independent Loop Free Alternatives – TI-LFA
In order to understand TI-LFA we need to first review
MPLS Fast ReRoute - FRR
IP Loop Free Alternatives - LFA
MPLS LFA

16. MPLS Fast ReRoute
FRR is a local detour method (not an end-to-end APS method)
to reroute quickly we pre-prepare labels for the bypass links 10 11 12
swap +
push
pop 13 14
from here on
no difference!
when link is down
change fwd table
protection LSP

17. LFA LFA FRR is another precomputed path Fast ReRoute mechanism
works with plain IP or MPLS
exploits an alternative to the default next hop
the alternative forwards to the destination (in any case)
without passing through the failed element (loop free)
A Loop Free Alternative
with respect to an element (link/node)
for a destination
is a router/LSR that
is not the default next hop
is connected to the destination
does not forward through the element
hence does not need to know about the failure) D S 1 2 3
default NH
LFA

18. Finding LFAs In order to a loop free alternative node
1. find the set P
all nodes still connected to the source in the event of failure
2. find the set Q
all the nodes that forward to the destination without failure
3. find the set PQ = the intersection of P and Q
4. choose a closest node in PQ
Note that PQ may be empty
LFA does not always succeed! 1 2 PQ S P Q
closest D 3 4 5

19. MPLS LFA with targeted LDP
After choosing the closest PQ node
the source LSR needs to push 2 labels on the label stack
top label to reach chosen LFA
label to reach destination LSR D after LFA pops the top label
But how does the source LSR know the label the LFA uses to reach D ?
it must open a targeted LDP session to find out! P PQ PQ PQ P PQ P 4 3 2 1 Q 5
LFA
LDP S D 2 1 Q Q

20. TI-LFA TI-LFA exploits MPLS SR
to avoid having to open targeted LDP sessions
The source LSR knows all the labels from the IGP
and can build the MPLS SR label stack accordingly
using any PQ node
This capability is Topology Independent in the sense that
a loop free backup path is found
irrespective of the topologies before and after the failure
Using deeper label stacks affords more flexibility
Immediately upon discovering the failure
the source LSR can use the new SR label stack
so the protection switch time is minimal